International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169

Volume: 5 Issue: 7 168 174

_______________________________________________________________________________________________
Statistical Feature based Blind Classifier for JPEG Image Splice Detection

Surbhi Gupta Neeraj Mohan

Research Scholar Assistant Professor
Computer Science & Engineering Computer Science & Engineering
I. K. Gujral Punjab Technical University I. K. Gujral Punjab Technical University
Kapurthala, India Kapurthala, India
royal_surbhi@yahoo.com erneerajmohan@gmail.com

AbstractDigital imaging, image forgery and its forensics have become an established field of research now days. Digital imaging is used to
enhance and restore images to make them more meaningful while image forgery is done to produce fake facts by tampering images. Digital
forensics is then required to examine the questioned images and classify them as authentic or tampered. This paper aims to design and
implement a blind classifier to classify original and spliced Joint Photographic Experts Group (JPEG) images. Classifier is based on statistical
features obtained by exploiting image compression artifacts which are extracted as Blocking Artifact Characteristics Matrix. The experimental
results have shown that the proposed classifier outperforms the existing one. It gives improved performance in terms of accuracy and area under
curve while classifying images. It supports .bmp and .tiff file formats and is fairly robust to noise.

Keywords-component: Blocking Artifact Characteristics Matrix (BACM); Image Forensics; Image Splicing; Joint Photographic Experts Group
(JPEG) compression artifacts; Support Vector Machine (SVM) classifier.

__________________________________________________*****_________________________________________________
I. INTRODUCTION
The readily available software, tools and techniques have
made the image processing quite easier these days. Tools
developed for enhancement of image are being misused to
hide the truth and establish the fallacies. There are enormous
ways to manipulate or forge an image. Most common image
forgery techniques are copy-move and splicing as shown in
Fig. 1. In copy move forgery, some part of the image is a)
cropped, processed and then replicated in the image to either
hide or add some content to the image. In splicing, two
different images are used to create a new image with new
content altogether. Thus, before relying on an image we
need to first check its truthfulness using image forensic tools
and techniques. These techniques are based on active and
passive approaches. In active approach, features like
watermark or signature is added to the image which would
get distorted if the image is tampered. This is mainly used b) c)
for sensitive documents and images, as they are highly Fig. 1 a) Original image; b) copy move forgery; c)
prone to fakery. In the absence of such active approach, a splicing forgery (Dong and Wang, 2011)
passive approach needs to be used. Passive approaches do
not require any background information about the image Initially, Lukas and Fridrich1, 2003 and Lukas et al.2, 2006
rather they extract features and characteristics from the proposed image tamper detection by identifying source
available image only to make a decision. camera using sensor pattern noise but it fails to correctly
Most of the image processing tools and digital cameras now classify the regions where the pattern noise was low. Ng and
days are using Joint Photographic Experts Group (JPEG) Chang3, 2004 proposed physics based model to detect image
format, so, the forensics for this format is very crucial. JPEG splicing but the detection rate was moderate. Popescu and
image forensics is done either by source or camera detection Farid4-6 (2004; 2005a; 2005b) presented image resampling
or by utilizing compression characteristics to identify image and color filter interpolation based methods to detect image
tampering. These characteristics are based on quantization splicing. Proposed method5 doesnt perform well where
and Discrete Cosine Transform (DCT) artifacts present in images with high quality factors were spliced and resaved at
the image due to double compression. a low quality factor. Pan et al.7 (2004) and Perra et al.8
(2005) utilized edge based features for detecting blocking
artifacts in JPEG images and achieved good results. Fan and
Queiroz9 (2003) introduced Blocking Artifact
Characteristics Matrix (BACM) based features to identify
168
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
International Journal on Recent and Innovation Trends in Computing and Communication
Volume: 5 Issue: 7
_______________________________________________________________________________________________
double image compression which Luo et al.10 (2007) used to
determine cropping and forgery, but this method gave a low
true positive rate. Chen and Hsu11(2008) investigated the
periodic property of blocking artifact by using different
features. But this method only performed well when forged
image has high quality factor as compared to original image.
Pan and Lyu12(2010) proposed region duplication detection
using image key-points and feature vectors as these are
robust to usual image transforms. Barni et al.13 (2010)
localized tampering by statistically analysing the image both
block and region wise. Bianchi and Piva14 (2012)
categorized the double JPEG compression as either aligned
or non-aligned and localized the tampering. Although results
presented were very comprehensive but classifier achieved
low Area Under Curve (AUC) for spliced images with high
Quality Factor. Thing et al.15 (2012) tried to improve the
accuracy of JPEG image tampering detection by considering
the characteristics of the random distribution of high value
bins in the DCT histograms. Then, Tralic et al.16(2012)
proposed a method to detect re-compression using Blocking
Artifact Grid extraction but sufficient illustration of method
on different types of images was lacking. Mall et al. 17(2013)
proposed a combined hashing index for image which was
capable of detecting structural tampering, brightness level
adjustment and contrast manipulations. Chang et al.18 (2013)
proposed copy move detection by searching similarity
blocks in the image and used similarity vector field to assure
the true positives. Recently, Wattanachote et al.19 (2015)
utilized BACM features to identify seam modifications in
JPEG images and presented efficient results.
All these researchers contributed significantly in image
forensics but only few provided a comprehensive study. The
aim of presented work is to design and implement a blind
classifier for splice detection of JPEG images at various
quality factors with higher accuracy and area under curve.
Proposed classifier works for .bmp and .tiff images as well.
It is robust to presence of noise in images. It detects image
splicing even when pre-processing and post-processing
operations have been applied and spliced area vary from
small to large. The proposed design and the experimental
results obtained are discussed in following sections.
II. PROPOSED SYSTEM DESIGN FOR SPLICE DETECTION
CLASSIFIER
The system design consists of two main components i.e.
training and testing of Support Vector Machine (LIBSVM 20)
to classify images as shown in Fig. 2. Image dataset
consists of original and spliced images from CASIA21
database. Dataset is divided as training and testing dataset.
Statistical features from these images are extracted from
image Blocking Artifact Characteristic Matrix (BACM)
which is the mean inter-pixel intensity difference inside and
across the JPEG sub-block boundaries. This difference is
similar for uncompressed images but when an image is
compressed, the discontinuities appear in pixel intensity
difference. The statistical features of images from training
dataset are fed to SVM and a model is obtained. Then this
model is used to test images for their identification as
original or spliced.
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
ISSN: 2321-8169
168 174
Fig. 2 System Design for proposed JPEG tool
A. Proposed algorithm for statistical features extraction
The algorithm used for extracting image statistical features
and its complexity is as follows:
Step1: Consider an image I. transform the image I to
grayscale such that Ig=rgb_to_gray (I).
Step2: Subdivide the image into sub-blocks of 8 x 8 pixels.
For each sub-block, for every pixel location , ,
where,1 , 8
Calculate difference in neighbour pixel intensities
(, ) as:
(, ) = |[(, ) + ( + 1, + 1)] [( + 1, ) +
(, + 1)] (1)
Where,(, ) represent intensity of pixel at location , .
Calculate ( + 4, + 4).
Calculate absolute difference (, ) = | + 4, +
4 (,)|. (2)
Step 3. Calculate energy (, ) at each pixel location
, from each sub-block as
(, ) = =1 (, ) (3)
Where, is total number of image sub-blocks.
Step 4. Calculate BACM matrix (, ) as (, ) =
(, )/. (4)
Step 5. Extract features F1-F20 from BACM and input them
to SVM to obtain the classifier model.
The algorithm works on 2x2 pixel neighbouring in each sub-
block. Every pixel is considered neighbour to 4 pixels as
shown in Fig 3. Algorithm needs to access each block once
and each pixel of the image 4 times to calculate pixel
intensity difference. So, the number of access for each pixel
is 4 and the complexity is equivalent to O (4n) O (n). It is
linearly dependent on the size of the image. The algorithms
main steps i.e. extracting BACM and defining feature set are
further clarified with example in the following two sections.
B. Extracting BACM
Blocking Artifact Characteristics Matrix (BACM) is a
matrix extracted form DCT blocks of the image. It reveals
important features about the image compression history. To
extract BACM, grey scale image is subdivided into sub-
blocks of 8x8 pixels. For each sub-block and every pixel
169
International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 7 168 174
_______________________________________________________________________________________________
location the inter-pixel intensity difference is calculated. For (1,1) = |(11 + 22 ) (21 + 12 )| and (5,5) =
example, P, Q, R and S are four consecutive sub-blocks in |(55 + 66 ) (65 + 56 )|
image. Then for sub-block P, the inter-pixel distance (4,4) = |(44 + 55 ) (54 + 45 )| and (8,8) =
at = = 1 , is calculated as (1,1) and 5,5 and the |(88 + 11 ) (81 + 18 )|
inter-pixel distance at = = 4 is calculated as (4,4) Further, the absolute difference (, ) is calculated using
and(8,8)using Eq. (1) as shown in Fig. 3. Eq. (2). Then energy (, ) and then BACM (, ) is
derived using Eq. (3) & (4). Fig. 4 shows the value of
BACM of an original JPEG image at each pixel location.
For example, 2.5364 in BACM is the mean value for all
pixels intensity differences which are located at (1, 1) in
every block.

a) b)
Fig. 3 Calculation of inter pixel difference a) inside b) across
the block boundary
2.5364 2.4623 2.5075 3.5274 2.6886 2.4883 2.6667 4.6139
2.5343 2.4259 2.3567 3.4163 2.3909 2.4273 2.5110 3.9047
2.3738 2.2702 2.4554 3.5679 2.4547 2.3628 2.3608 4.2888
2.5981 2.6049 2.8265 3.5741 2.7606 2.5171 2.6337 4.1214
2.7311 2.5343 2.8656 3.9444 2.9266 2.6317 2.7263 4.2798
2.5995 2.5816 2.6399 3.5178 2.5583 2.4067 2.5178 4.2305
2.6310 2.6879 2.9005 3.7634 2.7908 2.5583 2.7661 4.3909
3.3765 2.9156 3.0391 3.7558 3.2449 3.0192 3.4136 4.4266

Fig. 4 Sample BACM for JPEG image

BACM of an image gives important characteristics about it.
Experiments conducted on JPEG images at different quality
factors revealed that if an image with QF100 is spliced and
recompressed at same level the deviation in BACM values
increases as compared to original image. This deviation in
BACM values increases further if spliced images is
recompressed at lower levels as shown in Fig. 5.
18
16
14
BACM value
7x7matrix from BACM is considered for extracting features
but for proposed classifier whole 8x8matrix is considered.
Regions in BACM are defined as R1, R2, R3, R4, H1, H2,
V1, V2, C1, C2, and C3 and C4 as shown in Fig. 6. Further,
BACM is divided as R4, R5, R6, and R7 to extract
additional four features.
Spliced at QF
60

12 Spliced at QF
80
10 Fig. 6 Division of BACM in regions for extracting
8 Spliced at QF statistical features
100 The first set of features is based on symmetry of horizontal
6 region H1and vertical region V1. For H1 and V1
4 Original feature 1and 2 are extracted as:
1 = 3=1 | 4, 4,8 | (5)
2 3
2 = =1 | , 4 8 , 4 | (6)
1 2 3 4 5 6 7 8 Where (, ) represents BACM matrix value at
x value for y=4 location , . The next set of features is based on symmetry
of four regions R1, R2, R3 and R4. Feature 3 is based on
Fig. 5 Comparison of 4th column values of BACM of symmetry of R1 and R2, 4 is based on the symmetry of
Au_nat_00093.jpg with its spliced versions at QF100, blocks R3 and R4, 5 is based on the symmetry of blocks
QF80, and QF60 R1 and R3, 6 is based on the symmetry of blocks R2 and
R4, 7 is based on the symmetry of blocks R1 and R4 and
C. Defining Feature Set 8 is based on the symmetry of blocks R2 and R3.
3 = 3=1 3 =1 , (, 8 ) (7)
After calculating BACM, statistical features need to be
4 = 7=5 3 =1 , , 8 (8)
defined and extracted. For feature extraction, BACM is
divided in various regions. In existing techniques9, 10, 19, only 5 = 3=1 3 =1 , 8 , (9)

170
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 7 168 174
_______________________________________________________________________________________________
6 = 3=1 7 =5 , (8 , ) (10) Last two features19 and 20 are based on symmetry of
7 = 3=1 3 =1 , 8 , 8 (11) horizontal region H2 and vertical region V2:
8 = 3=1 7 =5 , 8 (8 , ) (12) 19 = 3=1 | 8, 8,8 | (23)
Further six features, 9 14 are extracted based on 20 = 3=1 | , 8 8 , 8 | (24)
percentage of occupancy of centre point C1 against different The values for all these features have been studied. Luo et
regions R1, R2, R3, R4, H1 and V1. These are calculated as: al., 2007 used first fourteen features i.e. 1 14 based on
9 = 1/ 3=1 3 =1 , (13) Eq. 5 to Eq. 18 to classify the images. In addition to these
fourteen features another set of six features based on Eq. 19
10 = 1/ 3=1 7=5 , (14)
to Eq. 24 have been added to increase the capability of the
11 = 1/ 7=5 3=1 (, ) (15) classifier. Another set of these features which are based on
7 7
12 = 1/ =5 =5 (, ) (16) the Occupancy of centre points C2, C3 and C4 have been
13 = 1/ 7 =1 4, 1 (17) studied but are not included in classifier design as less
14 = 1/ 7=1 , 4 1 (18) deviation is observed in their feature values. Fig. 7
Next four new features, 15 18 are extracted based on illustrates an example of feature values for original and
mean of four sub-regions i.e. R5, R6, R7 and R8 as: spliced images for existing and proposed classifiers for
15 = 4=1 4 =1 (, ) (19) image shown in Fig. 1. First fourteen (1-14) features are
16 = 8=5 4 =1 (, ) (20) common for both the classifiers and next six (15-20) are
4 8 added in the proposed classifier.
17 = =1 =5 (, ) (21)
18 = 8=5 8 =5 (, ) (22)

10
9
8 orig
Value obtained

7
6 tamp_
QF100
5
4 tamp_
QF80
3
2 tamp_
QF60
1
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Extracted Feature

Fig. 7 Representation of Feature values for original (Fig. 1a) and spliced images
(Fig. 1c at QF100, QF80, QF60)

III. EXPERIMENTAL RESULTS Different features studied are:

The experimental setup consists of images from CASIA
V2.0. 875 original and 665 spliced images are taken from
the database. Original and spliced images are saved at
different Quality factors 60, 80 and 100 to study the
classifier performance. Considered spliced images have
a. Randomly crop-and-paste image region(s)
b. Cropped image region(s) processed with resizing,
rotation or other distortion
c. post-processed region(s) (processed with operations
such as blurring) to finish crop-and-paste operation of
the fake image
d. Difference sizes (small and large) of spliced regions
e. Been considered at Quality factor 60, 80 and 100
f. Been considered to be realistic images by human eyes.
Training set consists of 900 images (both original at QF1
and spliced at QF2) and testing set consists of 640 images
(both original at QF1 and spliced at QF2). SVM classifier
with Radial Basis Function kernel is used. The penalty
parameter C is chosen by Grid Search method.
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
Accuracy and Area Under Curve for proposed
classifier
Accuracy of classification of images with small and
large spliced area
Impact of noise on classifier accuracy
A. Accuracy and Area Under Curve for proposed
classifier
The True Positive Rate (TPR), True Negative Rate (TNR),
Accuracy (ACC) and Area Under Curve(AUC) for the
existing and proposed classifier are compared in Table 1.
The TPR for existing classifier drops significantly when the
QF of spliced image is high but proposed classifier
maintained a promising TPR. TNR is almost comparable for
both the classifiers. Fig. 8 compares the overall accuracy for
both the classifiers. Accuracy for proposed classifier
remains high for all the scenarios. It is clear that the
proposed classifier outperforms the existing classifier in
terms of TPR and accuracy (ACC).
171
International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 7 168 174
_______________________________________________________________________________________________
Table 1 Performance comparison of Existing and Proposed Classifier

Original image Spliced Existing approach Proposed Approach

Quality factor image (Luo et al., 2007)
Quality
factor TPR TNR ACC AUC TPR TNR ACC AUC

QF100 QF60 95.4 98.6 97.0 0.9912 94.3 98.8 96.5 0.9971
QF80 71.0 95.4 83.2 0.8518 79.6 94.6 85.7 0.9195
QF100 71.0 96.0 80.8 0.8477 80.6 94.6 86.0 0.9067
QF80 QF60 86.7 98.5 92.4 0.9764 90.7 98.8 95.1 0.9796
QF80 74.1 95.0 85.2 0.8634 78.2 95.8 87.4 0.9168
QF100 70.3 97.1 83.8 0.8616 80.5 96.2 88.5 0.9324
QF60 QF60 66.1 96.3 82.2 0.8253 71.1 95.9 84.4 0.8905
QF80 81.7 96.9 89.8 0.9414 83.4 97.3 90.8 0.9712
QF100 86.5 97.5 92.6 0.9567 86.0 97.9 92.0 0.9639

100
98 Existing Proposed
96
94
92
90
Accuracy

88
86
84
82
80

Quality Factor of original/Quality factor of b)

spliced image Fig. 9 Receiver Operating Characteristic Curve for a)
existing and b) proposed classifier

A perfect classifier has AUC value equal to 1. Proposed

Fig. 8 Comparison of Accuracy for existing and classifier has AUC0.9 at all QFs. Moreover, the technique
proposed classifier achieved improved results without addition to algorithm
complexity.
Receiver Operating Characteristic Curve (ROC) for the
proposed and existing classifier is shown in Fig. 9. It can be
observed that the AUC values obtained for proposed
approach are much higher than those obtained for existing
classifier for all the scenarios.

B. ACCURACY OF CLASSIFICATION OF IMAGES WITH SMALL

VERSUS LARGE SPLICED AREA
The classifier performance is also evaluated in terms of
small versus large splicing area shown in Table 2. It is
observed that classifier performs better in classifying images
with small (<=30%) spliced area as compared to images
with large (30%-60%) spliced area as shown in graph in Fig.
a) 10.
Table 2 Performance comparison of proposed Classifier
for Small and large spliced area
Small Large
S. Quality
spliced spliced
No. factor
area area
1 QF60 98.9 98.8
172
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
International Journal on Recent and Innovation Trends in Computing and Communication
Volume: 5 Issue: 7
_______________________________________________________________________________________________
3 QF80 95.6 93.6
4 QF100 96.7 92.6
100 Small spliced area Large spliced area
99
Detection accuracy
98
97
96
95
94
93
92
QF60 QF80
Quality factor of spliced image
Fig. 10 Classifier accuracy for images with small and
large spliced area
C. IMPACT OF NOISE ON CLASSIFIER ACCURACY
As images are very much prone to noise, it is obvious that
the classifier results will vary in presence of noise. In this
paper four different types of noise are considered i.e. fast
fading, gaussian blur, white noise and JPEG. 320 images
with different types of noise from LIVE2 22database have
been taken. These authentic images are checked for their
true classification using the proposed classifier. The
classifier classifies the images having Gaussian blur and
white noise accurately. The accuracy obtained is 100%.For
Fast fading and JPEG noise, the accuracy decreases to
85.7% and 84.8% respectively. For more comprehensive
study, a number of original and spliced images with noise
may be tested. But it needs another experimental setup and
creation of new dataset by adding each type of noise to
various types of spliced images which is out of the scope of
this paper.
IV. DISCUSSION AND CONCLUSION
In this paper, machine learning based blind JPEG classifier
for detecting spliced images has been proposed and
implemented. The statistical differentiating features based
on image i.e. Blocking Artifact Characteristics Matrix
(BACM) have been extracted. Original images and spliced
images at various quality factors i.e. QF60, QF80, QF100
have been considered to train and test LIBSVM based
classifier. The main advantage of proposed classifier is that
it performs well irrespective of the quality factor at which
image is saved. It can be used to detect spliced images
undergone through any kind of pre-processing operation as
cropping, resampling, rotation etc. as well as any post-
processing operation such as blurring. Moreover, the spliced
area may be large or small. Additionally, it supports .bmp
and .tiff images. The receiver operating characteristic curve
and area under the curve demonstrated that proposed
classifier performs better as compared to existing one. The
only limitation is that classifier accuracy drops when both
the original and spliced images are saved at poor QF60.The
proposed classifier may be extended to make an integrated
forensic tool which can detect splicing, copy move, seam
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
ISSN: 2321-8169
168 174
carving, steganography and other types of tampering in
images.
References:
[1] J. Lukas and J. Fridrich: Estimation of primary quantization
matrix in double compressed jpeg images, Proc. of DFRWS,
Cleveland, USA, August 2003, pp. 5-8.
[2] J. Lukas, J. Fridrich, and M. Goljan: Detecting digital image
forgeries using sensor pattern noise, Proc. of SPIE, 2006, pp.
0Y10Y11.
[3] T.T. Ng and S.F. Chang: Blind detection of digital
photomontage using higher order statistics, IEEE
International Symposium on Circuits and Systems (ISCAS
QF100 2004), Vancouver, Canada, Vol. 5, pp. 688691.
[4] A.C. Popescu and H. Farid: Statistical tools for digital
forensics, International Workshop on Information Hiding,
Springer Berlin Heidelberg, 2004, pp. 128-147.
[5] A.C. Popescu and H. Farid: Exposing digital forgeries by
detecting traces of re-sampling, IEEE Trans. on Signal
Processing, 2005a, Vol. 53, no. 2, pp. 758767.
[6] A.C. Popescu and H. Farid: Exposing digital forgeries in
color filter array interpolated images, IEEE Trans. on Signal
Processing, 2005b, Vol. 53, no. 10, pp. 39483959.
[7] F. Pan, X. Lin, S. Rahardja, E. P. Ong and W. S. Lin :
Measuring blocking artifacts using edge direction
information, In Multimedia and Expo, ICME 2004, IEEE
International Conference on, Vol. 2, pp. 1491-1494.
[8] C. Perra, F. Massidda and D. D. Giusto: Image blockiness
evaluation based on sobel operator, IEEE International
Conference on Image Processing, 2005, Vol.1 (part-1), pp.
389392.
[9] Z. Fan and R. L. De Queiroz: Identification of bitmap
compression history: JPEG detection and quantizer
estimation, IEEE Transactions on Image Processing, 2003,
Vol. 12, no. 2, pp. 230235.
[10] W.Q. Luo, Z.H. Qu, J.W. Huang and G.P. Qiu: A novel
method for detecting cropped and recompressed image block,
in IEEE Int. Conf. on Acoustics, Speech and Signal
Processing, 2007, pp. II-217II-220.
[11] Y. L. Chen and C. T. Hsu: Image tampering detection by
blocking periodicity analysis in JPEG compressed images, in
Proc. of IEEE 10th Workshop on Multimedia Signal
Processing, Oct. 2008, pp. 803808.
[12] X. Pan and S. Lyu: Region duplication detection using image
feature matching,IEEE Transactions on Information
Forensics and Security,Vol. 5, no. 4, 2010, pp. 857867.
[13] M. Barni, A. Costanzo and L. Sabatini: Identification of cut
& paste tampering by means of double-JPEG detection and
image segmentation, in Proc. of ISCAS, 2010, pp. 1687
1690.
[14] T. Bianchi and A. Piva: Image forgery localization via block-
grained analysis of JPEG artifacts, in IEEE Trans.
Information Forensics and Security, Vol. 7, no. 3, 2012, pp.
10031017.
[15] V. L. L. Thing, Y. Chen and C. Cheh: An Improved Double
Compression Detection Method for JPEG Image Forensics,
IEEE Int. Symp. Multimedia, Dec. 2012, pp. 290297.
173
International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 7 168 174
_______________________________________________________________________________________________
[16] D. Tralic, J. Petrovic and S. Grgic: JPEG image tampering
detection using blocking artifacts, 19th International
Conference on Systems, Signals and Image Processing
(IWSSIP), IEEE, 2012, pp. 5-8.
[17] Mall, S. Shukla, S.K. Mitra and A.K. Roy: Comprehensive
image index and detection of tampering in a digital image, in
Int. Conf. Informatics, Electronics & Vision (ICIEV), 2013,
pp. 17.
[18] I.C. Chang, J.W. Yu, and C.C. Chang: A forgery detection
algorithm for exemplar-based inpainting images using multi-
region relation, Image and Vision Computing,2013, Vol. 31,
no. 1, pp. 5771.
[19] K. Wattanachote, T. K. Shih, W. L. Chang and H. H. Chang:
Tamper Detection of JPEG Image Due to Seam
Modifications, IEEE Transactions on Information Forensics
and Security, 2015, Vol. 10, no. 12, pp. 2477-2491.
[20] C.W. Hsu, C.C. Chang, and C.J. Lin: A practical guide to
support vector classification, 2003, pp.1-16,
http://www.csie.ntu.edu.tw/~cjlin/libsvm/.
[21] J. Dong and W. Wang: CASIA tampered image detection
evaluation database, 2011, http://forensics.idealtest.org.
[22] H.R. Sheikh, Z. Wang, L. Cormack and A.C. Bovik, LIVE
Image Quality Assessment Database Release 2, 2005,
http://live.ece.utexas.edu/research/quality.

174
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________